F5 patches five NGINX flaws that let outsiders run code remotely
The critical HTTP/2 bug, CVE-2026-123007, scores 9.2 out of 10 and can be chained with an unpatched kernel flaw toward root. Fixed builds are out, and no attacks in the wild have been reported.
F5 has published fixes for five vulnerabilities in NGINX that let an attacker anywhere on the internet run code with the rights of the NGINX worker, without signing in. CERT-EU said researchers have demonstrated CVE-2026-123007 and that no use of it in the wild has been reported. The server software, which CyberScoop said carries traffic for roughly a third of websites, should be updated quickly, with internet-facing systems first.
The critical HTTP/2 bug
F5 treats CVE-2026-123007, an integer underflow, as the worst of the set and rates it critical at 9.2 out of 10. The bug has been in the open-source code since version 1.9.5, shipped on September 22, 2015. Researchers call it nginx2root. Servers with HTTP/2 enabled are exposed.
Under HTTP/2 the client tells the server how much data it may send. When a client drops a request while the reply is still waiting, NGINX gives back the unused allowance and does not test whether the running total now passes the protocol cap. A later subtraction uses unsigned arithmetic, so the oversized total wraps to a huge value instead of falling below zero. That value is then taken as a negative length for the response, and a memory copy runs past the buffer.
Rapid7 researcher Stephen Fewer, named in the F5 advisory together with Wiz Research, Xint.io and Ikotas Labs, built an exploit against an NGINX 1.31.8 container with HTTP/2 and TLS enabled. The demo had to send roughly 128 gigabytes over HTTP/2, most of it control frames that carried no payload. It finished in two-and-a-half minutes on a fast local network. A 10-gigabit link would need about two minutes, and 1 gigabit would need 17 minutes. Because NGINX closes a connection after an hour by default, a link slower than about 300 megabits per second cannot complete the attack in time. The method also relies on a sped-up copy routine in the GNU C Library that is selected on x86-64 chips with certain AVX-512 features. Rapid7 said other methods may turn up, but none are known now.
The other flaws sit in different features: HTTP/2 and HTTP/3 reverse proxying, the resolver directive combined with mirror, and the mail proxy.
Root via the kernel, and what to install
Xint describes a separate local privilege bug in the kernel Stream Control Transmission Protocol code, introduced in the 4.15 kernel. Chained with the NGINX bug, it can give the attacker root. A kernel patch is not available yet. In a setup where NGINX ran in Docker, a root shell was obtained on the host. Many Linux distributions include the SCTP module, and the kernel loads it automatically once a program opens an SCTP socket. Whether that path is reachable still depends on the kernel, which modules are present, and the container’s restrictions. If it is reachable, the attacker can take the host kernel and, since containers share that kernel, every container on the machine.
Xint said it found the remote-code flaws using GPT 6.1 Sol and GPT-6 Luna. Wiz said it had found the same flaws on its own; a later account says Wiz Research, with Ikotas Labs, Metnir and Rapid7, independently found two of them.
The fixed releases are 1.31.9 and 1.30.7. Xint says operators who do not need SCTP should block that module until the kernel is fixed. CERT-EU tells administrators to install the NGINX update as soon as possible and to put internet-facing appliances first.